GitHub Stars Contributions MCP Server

Model Context Protocol server for the GitHub Stars program.

Current compatibility

• Package: current project version 0.3.1.
• MCP: official Python SDK 2.x (mcp>=2.1.1,<3), protocol revision 2026-07-28 with SDK compatibility for older clients.
• Contributions API: https://stars.github.com/api/contributions (REST).
• Profile/links/public Stars reads: GraphQL remains only where the reviewed migration does not document a REST replacement.

GitHub retired the old GraphQL contribution mutations on September 1, 2026. Contributions now use REST GET/POST/PUT. The REST API exposes no contribution DELETE.

Supported operations

Valid link platforms are TWITTER, MEDIUM, LINKEDIN, README, STACK_OVERFLOW, DEV_TO, MASTODON, and OTHER. Legacy aliases are accepted consistently: GITHUB -> README, WEBSITE -> OTHER.

Configuration

Stars/MCP settings:

• STARS_API_TOKEN — GitHub Stars token. Required for authenticated Stars reads/writes; discovery credentials never reuse it implicitly.
• STARS_API_URL — GraphQL base URL for remaining profile/link/public-profile operations; default https://api-stars.github.com/.
• STARS_CONTRIBUTIONS_API_URL — Contributions REST URL; default https://stars.github.com/api/contributions.
• STARS_AUTH_MODE — both|bearer|cookie; default both. REST normally needs bearer auth; both also preserves compatibility with remaining GraphQL calls.
• STARS_USER_AGENT — diagnostic User-Agent; default github-stars-contrib-mcp-server/0.3.1.
• LOG_LEVEL — DEBUG|INFO|WARNING|ERROR|CRITICAL; default INFO.
• MCP_TRANSPORT — stdio|http|streamable-http|sse; default stdio. http aliases streamable-http.
• MCP_HOST, MCP_PORT, MCP_PATH — HTTP bind configuration.
• VALIDATE_URLS — optional lightweight URL validation before Stars writes.

Discovery settings:

• DISCOVERY_DB_PATH — optional SQLite path. The default is the platform-local data directory ending in github-stars-contrib-mcp-server/discovery.db: %LOCALAPPDATA% on Windows, ~/Library/Application Support on macOS, and $XDG_DATA_HOME or ~/.local/share on Linux.
• GITHUB_DISCOVERY_TOKEN — optional dedicated GitHub REST token. Anonymous discovery is limited; this token is intentionally separate from STARS_API_TOKEN.
• YOUTUBE_API_KEY — optional YouTube Data API v3 key. Without it, only the limited public Atom-feed path for canonical channel IDs is available.
• DISCOVERY_FETCH_CONNECT_TIMEOUT_S, DISCOVERY_FETCH_READ_TIMEOUT_S, DISCOVERY_FETCH_MAX_BYTES, DISCOVERY_FETCH_MAX_REDIRECTS, and DISCOVERY_UNTRUSTED_EXCERPT_MAX_CHARS control the bounded safe-fetch surface.

The HTTP client retries transient 429 and 5xx responses up to three attempts with exponential jitter. Permanent 4xx errors are returned immediately. GraphQL enum errors for link platforms include the valid PlatformType values.

Contribution tools

• list_contributions(page=1)
• create_contribution(data)
• create_contributions(data)
• upsert_contribution(client_id, data)

The old GraphQL update_contribution(server_id, partial_data) contract is intentionally rejected: a legacy server-generated ID is not the same thing as the REST caller-controlled clientId and translating it can create duplicates.

Choose stable IDs such as talk:commit-conf-2026 or post:my-article-slug for repeatable idempotent writes.

Prompts and argument completion

Prompts are MCP prompts/* entries that validate a draft and return the tool call to perform. Validation happens before rendering, so a bad argument returns a structured error instead of a misleading plan.

A rejected argument returns a first line of [ERROR] <prompt name> followed by a JSON object with prompt, field_errors, an optional suggestions map (the valid ContributionType or group_by values) and an optional hint.

The server also registers a completion/complete handler, so clients autocomplete the type and group_by arguments by prefix.

Prompts are a core, non-deprecated MCP capability, but whether a given host surfaces them as user-invokable commands is host behaviour and is not claimed here. Roadmap placement and host-activation evidence are tracked in #96.

Autonomous discovery and review

The autonomous-discovery milestone adds a deterministic discovery/review pipeline behind MCP tools. The reusable orchestration layer lives in skills/ and agents/; it does not duplicate provider or publication logic.

Discovery/review tools:

• bootstrap_sources()
• list_sources(enabled_only=false)
• add_source(url, metadata, source_type)
• sync_source(source_id, dry_run=false)
• discover_contributions(source_ids, dry_run=false)
• list_candidates(states)
• get_candidate(candidate_id)
• review_candidate(candidate_id, decision, reason, edits)
• publish_approved_candidates(candidate_ids, dry_run=true)

Included skills:

• discover-my-contributions — bootstrap/sync trusted sources and build a review queue.
• sync-source — diagnose or refresh one source without provider bypasses.
• review-candidates — inspect evidence/provenance/conflicts and record explicit human decisions.
• publish-approved — mandatory dry-run first, then real publish only after explicit current user intent.

Skill delivery

The repository-root skills/* tree is the only authoring source for those workflows. It supports two delivery paths without copying the behavior:

• MCP Skills: clients implementing io.modelcontextprotocol/skills discover the skills with skills/list / skills/get and lazily read their manifest-authorized resources from this server.
• Standalone Agent Plugin: plugin.json packages this repository as github-stars-contributions; Agent Plugins v1 discovers the same root skills/ tree directly, so clients without MCP Skills support can consume the workflows without a second editable skill tree.

Agent Plugins 1.0 also supports packaging an MCP server through root mcp.json, but this repository intentionally does not ship that file yet. The published specification allows clients to omit/sanitize ambient subprocess environment variables, so inheriting STARS_API_TOKEN would not be a portable credential strategy. See docs/agent-plugin-mcp-packaging.md and #60 for the credential-safe bootstrap gate before plugin-bundled MCP registration is enabled.

APM consumers can also pin individual skills/<name> GitHub subtrees to an immutable Stars commit. See docs/skills-distribution.md for the single-source update model, APM example, central-catalog decision and duplicate/origin policy. Protocol/conformance evidence is in docs/mcp-skills-evidence.md.

Recommended user journey

1. Run bootstrap_sources() once to seed trusted source candidates from the existing Stars profile and contribution history.
2. Inspect the registry with list_sources() and explicitly add/verify only sources you control.
3. Run discover_contributions(...) or sync_source(...). Provider adapters emit neutral items and evidence; they never write Stars contributions directly.
4. Inspect list_candidates() / get_candidate(). Exact duplicates are blocked and ambiguous matches remain reviewable rather than being silently merged.
5. Record a human decision with review_candidate(...). High confidence alone never approves a candidate.
6. Call publish_approved_candidates(..., dry_run=true) first. A real write requires persisted approval, explicit current intent, and a fresh duplicate/policy check immediately before Stars REST.

See docs/sources/README.md for the source capability matrix and docs/security/discovery-threat-model.md for the security/privacy model. Evaluation methodology is documented in docs/evals/discovery-quality.md.

Host-neutral agent guidance is under agents/. Prefer MCP-served Skills when the connected host supports the extension; otherwise use the portable Agent Plugin or a pinned standalone skill import. Avoid activating both origins of the same logical skill in one session unless the host performs explicit origin-aware deduplication. Keep all secrets in environment variables or the host's secret store; do not copy credentials into skill or agent Markdown.

Deterministic server policy remains authoritative: fetched text is UNTRUSTED_SOURCE_CONTENT, high confidence never auto-approves, and publication only accepts already-approved candidates after a fresh Stars duplicate/policy check. X/LinkedIn scraping, browser-session reuse and cookie bypasses are not part of the workflow.

Local state and privacy

Discovery state is local SQLite: configured sources, cursors, candidates, bounded evidence, reviews, publication records and run summaries. Treat the database as personal workspace data and protect/delete it using normal local-file controls. Telemetry is deliberately content-free: structured discovery metrics/log fields may include run ID, source type, capability/status, counts, duplicate class, duration and bounded error class, but not titles, descriptions, page bodies, prompts, credentials or unnecessary URLs.

Running

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e '.[dev]'
cp env.template .env
export STARS_API_TOKEN='...'
python -m github_stars_contrib_mcp.server

For Streamable HTTP:

MCP_TRANSPORT=streamable-http MCP_PORT=8766 \
  python -m github_stars_contrib_mcp.server

Testing

Core release-quality checks:

python scripts/quality_check.py
pytest -q tests/evals/test_discovery_eval.py tests/evals/test_release_flow.py
pytest -q tests/unit/observability/test_discovery_observability.py
pytest -q tests/unit/test_server.py tests/unit/tools/test_discovery_tools.py
python -m compileall -q src
make test

Skill contracts and distribution invariants:

pytest -q tests/skills/test_skill_contracts.py tests/unit/skills

Stars API integration tests are isolated in a separate workflow and use STARS_API_TOKEN when configured. Mutation tests remain opt-in with STARS_E2E_MUTATE=1; they use stable PUT client IDs because REST does not provide DELETE cleanup. Release verification must not claim credentialed mutation evidence when no explicit token is available.

The source of truth for MCP tool schemas is src/github_stars_contrib_mcp/tools/.